Skip to content

[KERNEL][ROCm]Native HIP MXFP4(Compressed+Quark) (dense + MoE) for RDNA3 #46676

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
JartX wants to merge 41 commits into
base: main
Choose a base branch
from
Open

[KERNEL][ROCm]Native HIP MXFP4(Compressed+Quark) (dense + MoE) for RDNA3 #46676

Show file tree
Hide file tree
Changes from 33 commits
Commits
Show all changes
41 commits
Select commit Hold shift + click to select a range
1fc24bd
[ROCm][gfx1100] Add native HIP MXFP4 dense weight-only GEMM for RDNA3
JartX Jun 24, 2026
1f205e4
[ROCm][gfx1100] Add native HIP MXFP4 fused MoE for RDNA3
JartX Jun 24, 2026
b3446d7
[ROCm][gfx1100] Add native HIP MXFP4 MoE for GPT-OSS on RDNA3
JartX Jun 24, 2026
1d28923
[ROCm][gfx1100] MXFP4 dense: route batched decode (M>=2) through WMMA
JartX Jun 25, 2026
2631a71
[ROCm][gfx1100] MXFP4: scope kernel comments to gfx1100
JartX Jun 25, 2026
1f5f284
[ROCm][gfx1100] MXFP4 MoE: add WMMA path for batched decode / prefill
JartX Jun 25, 2026
2b034a4
[ROCm][gfx1100] MXFP4 MoE: occupancy-aware block_size_m
JartX Jun 25, 2026
307d76a
[ROCm][gfx1100] MXFP4 dense: LUT dequant for the scalar decode GEMV
JartX Jun 25, 2026
5c71b94
[ROCm][gfx1100] MXFP4 MoE: LUT dequant for the scalar expert GEMV
JartX Jun 25, 2026
3f7ef40
[ROCm][gfx1100] MXFP4: LUT dequant for WMMA path + raise scalar dispatch
JartX Jun 25, 2026
23722a3
[ROCm][gfx1100] MXFP4: move GPT-OSS gfx1100 selection out of generic …
JartX Jun 25, 2026
b16e783
[ROCm][gfx1100] MXFP4: apply pre-commit formatting (ruff + clang-format)
JartX Jun 25, 2026
6b0a024
[ROCm][gfx1100] MXFP4 MoE: rename tpe -> tokens_per_expert (typos hook)
JartX Jun 25, 2026
faf7db9
[ROCm][gfx1100] MXFP4 dense: gate linear kernel on gfx1100, not gfx11xx
JartX Jun 25, 2026
df10623
[ROCm][gfx1100] MXFP4: don't route W4A16 MoE through the WMMA-16 tile
JartX Jun 25, 2026
bfe3c81
[ROCm][gfx1100] MXFP4: test the W4A16 MoE tile-cap regression
JartX Jun 25, 2026
7fd46d3
[ROCm][gfx1100] MXFP4: pre-commit fixups (ruff-format + typos)
JartX Jun 25, 2026
d43f143
[ROCm][gfx1100] MXFP4 MoE: register an oracle backend instead of besp…
JartX Jun 26, 2026
c8db136
[ROCm][gfx1100] MXFP4 MoE: leave the W4A16 INT4 path untouched
JartX Jun 26, 2026
95f2626
[ROCm][gfx1100] MXFP4 MoE: ruff-format the experts backend
JartX Jun 26, 2026
2afb7c0
[ROCm][gfx1100] MXFP4 MoE: fix tensor-parallel reduce + route Quark/W…
JartX Jun 26, 2026
2495c06
[ROCm][gfx1100] Quark MXFP4: run weight-only dense linears on the nat…
JartX Jun 26, 2026
09ebe42
[ROCm][gfx1100] MXFP4 MoE: select RDNA3 via the priority list, not th…
JartX Jun 26, 2026
1608b99
[ROCm][gfx1100] MXFP4 MoE: trim verbose comments (no functional change)
JartX Jun 26, 2026
abc2f00
Merge branch 'main' into feat/rdna3_mxfp4_native
JartX Aug 13, 2026
daadfee
Register the RDNA3 MXFP4 linear kernel in the ROCm priority list
JartX Aug 13, 2026
6e5392e
Make the RDNA3 MXFP4 linear kernel follow the weight-only contract
JartX Aug 13, 2026
7105001
Drop the redundant MXFP4 op-availability guards
JartX Aug 13, 2026
f92711a
Take the MXFP4 MoE gate/up layout as a parameter, not from the config
JartX Aug 13, 2026
47495e2
Route the compressed-tensors MXFP4 MoE repack through the oracle
JartX Aug 13, 2026
bdd3681
Merge branch 'main' into feat/rdna3_mxfp4_native
JartX Sep 6, 2026
fd7833e
[ROCm][MXFP4] Make the RDNA3 MXFP4 experts follow the modular-kernel …
JartX Sep 6, 2026
10bb691
[ROCm][MXFP4] Bound the RDNA3 MXFP4 GEMM test by the rounding it can'…
JartX Sep 6, 2026
a2b7ae5
Merge branch 'main' into feat/rdna3_mxfp4_native
JartX Oct 6, 2026
224c825
[ROCm][MXFP4] RDNA3 MoE: align sizes to the MXFP4 block, not 256
JartX Oct 6, 2026
ff32da4
[ROCm][MXFP4] Drop the blank line after the converter docstring (ruff)
JartX Oct 6, 2026
e40a0ae
Merge remote-tracking branch 'upstream/main' into feat/rdna3_mxfp4_na…
JartX Oct 22, 2026
e060ed0
[ROCm][MXFP4] Address review: fix fused down-GEMM rows, follow the or…
JartX Oct 22, 2026
69ceda1
[ROCm][MXFP4] Make the experts test biases non-optional (mypy)
JartX Oct 22, 2026
91f53ed
[ROCm][MXFP4] Correct the down-GEMM reduction comment, document the R…
JartX Oct 22, 2026
ebb17e4
Merge remote-tracking branch 'upstream/main' into feat/rdna3_mxfp4_na…
JartX Oct 24, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 3 additions & 1 deletion CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1518,7 +1518,9 @@ if(VLLM_GPU_LANG STREQUAL "HIP")
list(APPEND VLLM_ROCM_EXT_SRC
"csrc/rocm/q_gemm_rdna3.cu"
"csrc/rocm/q_gemm_rdna3_wmma.cu"
"csrc/rocm/moe_q_gemm_rdna3.cu")
"csrc/rocm/mxfp4_gemm_rdna3.cu"
"csrc/rocm/moe_q_gemm_rdna3.cu"
"csrc/rocm/moe_mxfp4_gemm_rdna3.cu")
endif()

define_extension_target(
Expand Down
Loading
Loading